Exploiting High-Level Knowledge Resources for SpeechRecognition

Bücher

Angebote / Angebote:

This book proposes a novel methodology to improve theperformance of a Large Vocabulary Continuous SpeechRecognizer (LVCSR) by modeling several high-levelknowledge resources into an n-best list re-rankingmechanism. The book focuses on the identification andformulation of several novel, additional, domain-independent knowledge resources into are-ranking mechanism. We illustrate the extent ofimprovements obtainable by efficiently exploitingphonetic, lexical, syntactic and semantic knowledge.We improve WER for specific domains by combiningdomain-independent knowledge with automaticallyextractable domain-dependent resources. To modeldomain-dependent knowledge, we propose a methodologyto automatically generate SLMs for specific dialogstates. The heart of this book not only lies in thetask of selecting and modeling key informationresources but also on combining them efficiently.Hence, we explore using minimum error rate trainingto optimally assign knowledge resource weights bydirectly minimizing the WER on a development set.Finally, we present a novel IVR grammarcreation/tuning application and illustrate theimportance of the re-ranking mechanism in this framework.